Papers / Updated at: 2026-11-01

Das et al. (2025): Synapse: Adaptive Arbitration of Complementary Expertise in Time Series Foundational Models

Authors:

• Sarkar Snigdha Sarathi Das
• Palash Goyal
• Mihir Parmar
• Yiwen Song
• Long T. Le
• Lesly Miculicich
• Jinsung Yoon
• Rui Zhang
• Hamid Palangi
• Tomas Pfister

Source: https://arxiv.org/abs/2511.05460

DOI: https://doi.org/10.48550/arXiv.2511.05460

Mentioned in:

• Ensembles Are Called Agents Now
• Ahamed et al. (2026): Reasoning-Aware Training for Time Series Forecasting
• Das et al. (2025): Synapse: Adaptive Arbitration of Complementary Expertise in Time Series Foundational Models

I came across Synapse as the “STRIDE w/ Synapse” submission ranks first on the GIFT-eval leaderboard at time of writing.

A Synapse-only submission is available, too, and ranks at nearly identical performance to Toto-2.0-1B, Chronos-2, and t0-beta.

What stands out is that in terms of both MASE and CRPS the “STRIDE w/ Synapse” submission is noticeably better than all other submissions (second to only EXAONE-Forecast-Agent). Comparing the different combinations of STRIDE and Synapse, does the improvement stem from STRIDE or Synapse?

A complicating factor in assessing this without replication code is the fact that Synapse is an ensemble of foundation models. There is no way to tell whether the set of foundation models used for the two Synapse-based submissions was the same. The Synapse paper uses fairly old foundation models. Perhaps the “STRIDE w/ Synapse” submission used more recent foundation models (such as Chronos-2).

Overall, Synapse looks like a usual ensembling approach based on past model performance. What stood out to me as unique was the “Dynamic Weight Adaptation via Forward Sampling”, where Synapse adjusts the weights assigned to the constituent models over the forecast horizon by continuing to evaluate models—not against the unavailable actuals, but against the resulting ensemble prediction.

Wouldn’t this imply that, over the forecast horizon, weights converge to the model that started out with the largest weight? As the simulated ground truth is largely that model’s prediction, and so eventually prediction equals simulated ground truth?

If weights based on past performance assigned 100% of weight to one of the constituent models, its prediction would be the ground truth and the weights would not adjust. Or they would adjust because other models happen to predict the same value.

While the authors compare the hypothetically possible oracle MASE/CRPS at the beginning of the paper (comparing against the constituent models, only), they unfortunately do not include the oracle performance in the final results, avoiding an evaluation of Synapse based on how much of the theoretical improvement it achieved.

Alas, when including the oracle forecast, one gets the typical picture that ensembles tend to achieve little of the hypothetical improvement. When comparing with the difference achieved by the oracle, it becomes questionable Synapse truly “substantially outperforms” even just Moirai2.

Oracle values from Figure 2. Synapse and benchmark ensembles from Table 1. Rest available in both Figure 2 and Table 1.

Note that the values reported for Synapse in this table are worse than the ones currently reported on GIFT-eval (table above). In contrast, the values the paper reports for Toto and Moirai2 match the ones reported on GIFT-eval. No clue where that difference comes from (again, no replication code available). But: The difference between the Synapse MASE result reported in the paper and the MASE result on GIFT-eval (0.719 - 0.699 = 0.020) is larger than the improvement the paper reported for Synapse over Moirai2 (0.728 - 0.719 = 0.009). Take from that what you will.¹